AUTOMATED CLASSIFICATION OF MARMOSET VOCALIZATIONS AND THEIR REPRESENTATIONS IN THE AUDITORY CORTEX

Communication sounds are a subset of sounds that are characterized by their unique acoustic structures, biological relevance and specificity for social interaction. While there is difference between the complexity of human language and animal vocalizations, animals have been found to share similar vocal behaviors such as representative signaling and categorical perception with human. In addition, cortical organizations are similar between human and non-human primates. Therefore, it is believed that revealing the neural basis of primate vocalizations can give us insight on how human perceive speech. Over the past decades, several experiments have been conducted to elucidate the computational principles of neural representations of species-specific vocalizations in the auditory cortex of non-human primates. However, no consistent picture has emerged. Whether vocalizations are encoded by highly specialized cortical neurons through a specific pathway is still debated. This is partly due to the lack of consideration of statistical structures of vocalizations and the behavioral relevance of their acoustic features. In this thesis, I developed an automatic call classification and feature measurement algorithm to analyze a large number of marmoset vocalization recordings from a captive colony maintained by Dr. Xiaoqin Wang’s laboratory at Johns Hopkins University since 1996. These vocalizations were recorded from two groups of previous unfamiliar monkeys that were later placed in the same colony room (refer to hereafter as merging). We found that the call features of the current colony members (refer to hereafter as the current group) had similar distributions to those of the former members of this colony measured 18 years ago (refer to hereafter as the previous group). It is also found that after merging, the dynamics of convergence/divergence were observed in their vocalization acoustic features. Besides acoustic features, we also examined the percentage of different call types in each session and the temporal sequences between and within the two groups of marmosets. Call type percentage was found to be different between the current and previous groups. While marmoset Trill calls were found to occur more frequently than Twitter calls in the previous group, Twitter calls occurred more often than Trill calls in the current group. The call type percentage of the current group remained relatively stable across seven months sampling period. It is found that temporal correlation between the call timing increased during the formation of a new social environment, and went back to the average a few months later. These findings suggest possible vocal plasticity and how temporal structures could vary under different social situations. In the second part of my thesis, I used a synesthetic vocalization stimulus design to systematically probe the neural coding of vocalizations in marmoset auditory cortex with consideration of statistical structures. We found that A1 neurons showed relatively low selectivity to vocalizations and the selectivity increased in lateral belt neurons. Receptive field of these neurons also indicated vocalization features were processed with a distributed code in A1. How spatial location and background noise affected neuronal selectivity and tolerance of vocalizations was also explored. Overall, this thesis focuses on a new approach to study monkey vocalization structures, their communication behaviors, and computational principles of vocalization processing in the auditory cortex. Preliminary data was presented in neural recordings